How AI Agents Are Revolutionizing E-commerce — Lessons from Real-World Deployment

Summary

Many teams claim to build AI agents, but few have the experience of directly conversing with real customers and driving purchases. WADDLE is one such team, having successfully applied their AI salesperson 'Gentoo' to real-world e-commerce, earning recognition from OpenAI for their technology and execution, leading to a collaboration.

In this [🎙️KV Talk Plus+] episode, we sit down with CSO Yongwon Cho of Waddle to explore:

• How AI agents are actually working in commerce today

• What execution, UX, and strategy make it happen

• The 4 key success factors we at Kakao Ventures look for in AI agents

Chapters

00:00 Intro
00:54 Gentoo, the AI Salesperson for Online Stores 🛍️
01:41 How Sam Altman Defines “AI Agents”
02:34 Why Commerce Is the Perfect Playground for AI Agents
04:33 The Real Game: Personalization & Connectivity
06:15 Gentoo Seemed Like a Genius… But
07:36 '4 Competitive Advantages' for AI Providers ✨
08:27 How Gentoo's AI UX Was Designed
09:39 Conversational Interface: The Secret to a 2x Conversion Rate Boost
11:06 Collaborating with OpenAI: What Made the Difference
13:36 WADDLE's Insights Gained from Global Play
14:22 Waddle’s Playbook for SaaS Success
16:23 How Waddle Cracked B2B Sales
18:25 The Core Belief That Guided Waddle Through Everything
20:19 For Gentoo, the real competition is Siri.
22:25 Outro

Transcript

Youngmoo Kim: Hello, I'm Youngmoo Kim, an Associate at Kakao Ventures, in charge of deep tech investments. In the last episode, we discussed how an AI agent is technically created and what real-life examples could emerge.

In this episode, we've invited Yongwon Cho, CSO of WADDLE, a Kakao Ventures portfolio company, to discuss how AI agents are being developed and utilized in the real market. We'll be talking about how AI agents are being developed and utilized in the real market. Here's Yongwon Cho, CSO of WADDLE.

Yongwon Cho: Hello. I'm Yongwon Cho, Co-founder of WADDLE. We are developing 'Gentoo,' an AI salesperson solution that converses with e-commerce users and recommends customized products.

Gentoo, the AI Salesperson for Online Stores

Youngmoo Kim: What exactly does Gentoo, an agent specialized in e-commerce, primarily do?

Yongwon Cho: When you go to an offline store, there's a salesperson, right? They greet customers, answer questions about products, recommend products, and perform various other functions. Online stores are currently closer to "unmanned stores." In these unmanned stores, an agent named Gentoo interacts with users, converses with them to understand what products they are looking for and why, recommends customized products, and thereby increases the purchase conversion rate.

How Sam Altman Defines “AI Agents”

Youngmoo Kim: Before we delve into our discussion, I'm curious about what WADDLE defines as an AI agent, and why it's necessary to utilize AI agents specifically in e-commerce.

Yongwon Cho: The concept of an AI agent is still in the process of being established. So, rather than me saying, 'This is what an agent is,' I often borrow definitions from well-known figures.

Last October, at the Dev Day held in London, which OpenAI ran privately, someone asked Sam Altman, 'What is an AI agent?' Sam Altman's response was that it's a system that autonomously solves complex tasks with only high-level instructions. Complex means not a problem with a fixed answer or a fixed workflow, but one with various factors and variables. The AI decides which tools and data to use to solve this problem, and this process is autonomous without human intervention, which is a characteristic of agents.

Why Commerce Is the Perfect Playground for AI Agents

Based on this definition, when we talk about why AI agents are necessary for e-commerce, E-commerce, among many industries, is particularly an industry fraught with complex problems. Just looking at the user experience initially, users need to encounter products when they enter a shopping mall, make purchase decisions for those products, obtain information, eventually make a payment, receive delivery, and there are scenarios corresponding to delivery confirmation. And each of these stages is interconnected, making it a very complex industry. The goal of all e-commerce companies is uniformly to increase sales and reduce costs.

However, within this, a wide variety of strategies are formulated. You can increase repurchase rates, increase conversion rates, or issue coupons. There are various strategies like these. These aren't always concrete strategies; rather, high-level strategies like "this is how we do it" are established. We started applying AI agents to e-commerce with the insight that the various complex problems in e-commerce can be solved most effectively through AI agents.

Youngmoo Kim: Indeed, from a customer's perspective, they often enter a shopping mall with a vague idea like, "I'm in this situation, is there a product for it?" rather than needing a specific product, so it definitely seems like an AI that can autonomously solve those scattered problems at once would be beneficial. That was probably the role of a salesperson in offline stores, and Gentoo's role is to make that possible online, right?

Yongwon Cho: Yes, that's right. Simple AI chatbots that answer A when asked A have been effective so far, but that's not it. Instead, some agents constantly analyze user intent, some are good at recommending, and some formulate strategies to encourage users to repurchase. We believe that online commerce is an industry where a combination of agents, each best at their respective tasks, can perform much better than existing chatbots or humans, with user information constantly being communicated between these agents.

The Real Game: Personalization & Connectivity

Youngmoo Kim: For Gentoo, the AI service doesn't exist as its own site like ChatGPT, but rather is integrated into the customer's website. Because of this, even though customization and connection, which were explained in the previous episode, might be difficult and challenging areas, you are taking on the challenge in the e-commerce domain. So, from that perspective, how does WADDLE view the importance of these two factors?

Yongwon Cho: What we first focused on was the “conversational interface” itself when interacting with AI. Compared to the past where we entered commands, clicked, or performed simple actions on computers, as people started conversing with ChatGPT, actual human intentions and qualitative data could be connected with AI. We began to think that we could solve real-world problems with more data and more connection points by using this AI technology. That's when we started attempting to apply AI to e-commerce, which is the business area we've been in for the longest time.

We found that while a lot of data is accumulating, such as product data and user preferences or intentions that users express when they come in, like why they buy or don't buy, companies are unable to utilize this data because they don't have many data engineers or well-equipped data analysis tools. Conversely, from the consumer's perspective, we discovered that there are problems where users clearly want to purchase a product and it exists on the site, but they just can't find it. WADDLE is solving this problem through an AI salesperson.

Youngmoo Kim: Since we need to respond to and persuade customers who are in the purchase journey, “customization” was necessary, and because we need to connect with the existing workflows of e-commerce companies, “connection” is important. You probably had many trials and errors in developing the product in this way. What kind of challenges did you mainly encounter?

Gentoo Seemed Like a Genius… But

Yongwon Cho: I can give you an honest answer. When we first created Gentoo and had internal QA sessions talking with Gentoo, the first thought that came to mind was, "Our child must be a genius." Even just by connecting product data and review data well, the accuracy of recommendations and conversations was quite high from the beginning.

However, when we actually deployed this AI agent Gentoo to e-commerce shopping malls and served actual users, we started seeing results that were slightly different from what we had expected. Because the user experience of Gentoo changes from a shopping mall frequently used by men in their 30s to one used by women in their 20s.

Initially, we thought that good recommendations and good conversations would immediately lead to purchases, but for each shopping mall, or for each product within a shopping mall, or for each user of a shopping mall, we are continuously accumulating data on the appropriate purchase journey, when Gentoo should appear, what kind of conversations should happen at what time, and what questions should be asked, and based on that, Gentoo's conversations and recommendations are constantly changing.

4 Competitive Advantages for AI Providers

Youngmoo Kim: It must have been a challenging product development, yet you developed Gentoo quite rapidly. So, when we were reviewing the investment, we often wondered what kind of competitiveness WADDLE could gain in terms of product structure and the challenges encountered while meeting customers. Kakao Ventures believed that AI agents, unlike general LLM wrappers, could gain competitiveness in four aspects.

First, it was 'AI Experience Design,' which looked at how well the product was designed UI-wise based on users utilizing AI. Second, it was 'Ecosystem Synergy'. AI agents are a new technology, and it seems like a market where we shouldn't build this technology alone, but rather in cooperation with all players in the value chain related to that technology. We considered how to create synergy in such aspects.

Third, to enable expansion, we need to successfully develop the product into a SaaS solution, so how can we package it effectively? Lastly, how can we effectively sell this new type of product, an agent? These were the four perspectives. WADDLE seems to have a clear strength in each of these four aspects, especially regarding AI User Experience, which was the first point. How did you design and develop it so far?

How Gentoo's AI UX Was Designed

Yongwon Cho: Technologically, Gentoo possesses the characteristics of an AI agent architecture and an app, but it can also be classified as a 'conversational interface' solution. This means that users conversing with AI is a very significant part of the Gentoo experience. That's why we are currently dedicating the most resources to designing user experiences specialized for e-commerce.

At what timing during the shopping process would a user need a salesperson? There might be users who don't know how to search when they first enter, or users who want to compare products after searching, wondering which one to buy between two items. There could also be users who lack enough information to make a purchase decision, or users who need price assurance.

We are continuously experimenting with these diverse situations, breaking them down to determine when Gentoo should appear, what kind of conversations it should have, and what the results are, thereby designing the Gentoo user experience.

Conversational Interface:
The Secret to a 2x Conversion Rate Boost

Youngmoo Kim: Indeed, given it's an AI chatbot, chat isn't just about replying; it needs to lead to actions like actual recommendations. I imagine there were concerns related to this, and also regarding the interface itself, since Gentoo presents a new form of UI compared to the image or button-based UIs we're used to, you must have had many concerns. Have there been any breakthroughs in this area recently?

Yongwon Cho: From a practical standpoint, the most insightful experiment was, though it sounds obvious, the user experience begins with Gentoo appearing and initiating a conversation with the user. If you think of it as a salesperson in an offline store, they can't just stand still, can they? There needs to be a trigger for them to speak to the customer when they enter.

This trigger was initially applied uniformly to all pages. It was a very simple flow where it would greet you whether you came from the home page or a product detail page. One of the recent changes was to design different greeting messages for Gentoo based on the current page or the user's actions on that page. For example, if a user is holding a product, Gentoo should say, "That product is great. Here's why." If the user is Browse back and forth, Gentoo should offer help to prevent them from getting lost.

We recently tested having Gentoo initiate conversations with these kinds of phrases. As a result, we actually saw a 2-3x increase in conversion rates. So, from now on, Gentoo and our engineers will work together to design in detail what scenarios Gentoo should initiate conversations in.

Collaborating with OpenAI: What Made the Difference

Youngmoo Kim: It seems you've reached a product level where simply connecting an LLM model to an API wouldn't easily achieve such an impact. In fact, the reason we could do this so quickly is because, like you, CEO of WADDLE frequently attends conferences in the US and overseas, gathering a lot of information. I heard you also attended ICLR last week.

And WADDLE, in particular, is the first company in Korea to officially sign a contract with OpenAI. Could you give us an example of how bringing in external synergy and playing directly has truly been beneficial?

Yongwon Cho: We are a company that excels at building applications and designing user experiences using LLMs, and ultimately, our task is to implement AI agents effectively. We are continuously and dynamically dividing the problems we need to solve, the problems we can solve, and the problems that can be solved according to technological trends, and we are operating our team based on that.

Among these, what we received significant help from, and at the same time, desperately wanted to succeed in, was our connection and cooperation with OpenAI, and our enterprise agreement. We visited OpenAI once in March and once in October last year, met their internal engineers and GTM team, and successfully secured an enterprise agreement.

OpenAI also invited us to several closed events, telling us that they would focus on solving certain problems, so companies wouldn't have to worry about those aspects. Conversely, they also mentioned that some problems have very regional or industrial specificities, and they say it's right for startups to solve these problems. In fact, the reason we were "chosen" by OpenAI, so to speak, was because WADDLE was creating real-world examples of AI agents and conversational data being utilized in the Korean e-commerce market.

And as you mentioned, what we excel at is not really building the underlying models, improving conversational accuracy, or increasing conversational speed, but rather things that are closer to the actual user. Therefore, we have a direct channel to communicate with the team that is building these language models themselves, and we receive their assistance in designing the architecture between models, such as how to best use these models and how to best integrate them for maximum effectiveness.

WADDLE's Insights Gained from Global Play

Youngmoo Kim: While you spoke very humbly, what I felt during our monthly meetings since investing in WADDLE was, 'This is truly a global team.' In fact, I've witnessed firsthand Korean startups honestly collaborating with American startups that are building models and tracking the fastest AI trends in the US, discussing how they can research and develop AI together. I definitely felt that WADDLE is not only doing great with OpenAI but also leveraging the developer community exceptionally well, truly operating as a global team.

Yongwon Cho: We are also continuously interacting with companies, especially in Silicon Valley and New York, to build their product. They are also continuously interacting with companies, especially in Silicon Valley and New York, to build their product. These companies have two main advantages: Because they are companies building products in the fastest technological current, overcoming high competitive intensity, the performance of their solutions is improving at a very rapid pace. Conversely, since we are an AI agent company with actual corporate clients in Korea, we provide them with real-world examples, and as they further refine their products, they sometimes solve problems that we couldn't, creating a continuous synergy. The second point, from a more mental aspect, is that experiencing their development speed and the global market trends firsthand continuously motivates our team. That's excellent.

Waddle’s Playbook for SaaS Success

Youngmoo Kim: Ultimately, You succeeded in making an AI product into a SaaS product. There must have been a lot of struggle in that aspect, too. How did WADDLE manage to navigate through it well?

Yongwon Cho: One of the characteristics of Gentoo as a SaaS is that while Gentoo converses with consumers(users), the decision to adopt Gentoo is made by the company's representative. This is sometimes referred to as B2B2C. We focused on two points.

Although this is essentially a B2B bible, the first objective was to make the adoption process easy by minimizing friction and reducing resources as much as possible. And the second was to enable the person in charge to create results by using Gentoo and become a hero within their company. We are focusing on these two aspects for our SaaS productization.

Regarding the first point, making adoption easy, we have almost automated the process of integrating shopping mall data, training Gentoo, and deploying Gentoo as an SDK to shopping malls. If a company has developed its own shopping mall, they can integrate their own shopping mall DB and install the SDK to use it. Companies using hosting solutions can simply click, and Gentoo can appear on their shopping mall screen automatically within about a day. This is the first advantage: confirming Gentoo's performance and adopting it quickly without various procedures.

On the other hand, ultimately, the person in charge must become a hero internally through Gentoo for them to continue using Gentoo and become its champion within the company. To achieve this, we are transparently showing the results Gentoo is producing, such as how many people have interacted with it, what conversations they had, what they frequently asked, and how many made purchases.

Furthermore, we recently released an update that allows company representatives to directly coach Gentoo, by providing it with guidance, just like entering prompts into ChatGPT, to make Gentoo an even better salesperson for their shopping mall.

Youngmoo Kim: Many startups fail at product SaaS-ification and end up reverting to SI(System Integration), so I believe it's crucial to find that middle ground, and this team has successfully found that point.

Speaking of the B2B aspect in B2B2C, productization ultimately relies on smooth sales. So, selling this new type of product, an agent, to managers must also be a new challenge. Were there any instances where WADDLE's sales approach proved particularly effective?

How Waddle Cracked B2B Sales

Yongwon Cho: When I first introduced Gentoo, I often used the analogy of an 'offline store salesperson,' and I really like this analogy because when I used it, every person in charge of an e-commerce shopping mall understood it immediately.

The concept is clear:

• Offline stores have salespeople

• But online stores are unmanned

• So Gentoo can help you

This flow is clear. However, while it's very easy to convince people up to this point, the first hurdle was, "What kind of conversations can Gentoo actually have?" And the second was proving to the person in charge what kind of results they would see from actually conversing with Gentoo. These two points always had to be demonstrated during sales.

Conversely, if we prove these points, sales become easier, and the sales conversion rate increases. So, what we did was allow Gentoo to learn from product data within about a day, enabling the person in charge to actually use Gentoo within the console. This eliminated many questions. By conversing with Gentoo, they could see what it could do, what it couldn't, and where data was lacking.

And by adding features that allow managers to coach Gentoo, just like inputting prompts into ChatGPT, they could directly incorporate their own shopping mall operating expertise into the areas where conversation was lacking. This was a method that solved the managers' concerns and made sales easier. Secondly, as Gentoo integrates, consumers actually make purchases based on Gentoo's recommendations, and they also ask questions and make purchases through Gentoo.

As this data continuously accumulates and we gain various insights by analyzing this conversational data, sales are improving as we quantitatively and qualitatively demonstrate that Gentoo is performing well through case studies and conversational data analysis.

The Core Belief That Guided Waddle Through Everything

Youngmoo Kim: Listening to all this, you explain product development so smoothly that it might seem easy, but we know it's incredibly difficult. Therefore, if there was a particularly memorable difficult moment while operating WADDLE, what was it, and how did WADDLE overcome it in its unique way?

Yongwon Cho: WADDLE was originally a company that operated a conversational shopping mall for visually impaired people, and from that, it was a B2C platform company that operated a shopping mall using actual conversational data from visually impaired people.

One of the consistent beliefs we've held since then is that the world should move towards more intuitive interfaces that more people can use. We've been continuously developing products with the intuition and belief that 'conversation' will be the most intuitive interface when looking 5 to 10 years into the future.

Then, when ChatGPT emerged, many people started conversing with AI. They might not have realized it, but they went from touching screens to suddenly typing into ChatGPT, especially with the GPT app, and conversing by voice. I feel these are very significant changes.

So, the moment this 'conversational' interface started to be perceived as more intuitive by many people, I felt a great thrill, thinking that our long-held belief could actually be realized. And once we actually supplied that product to B2B shopping malls, and real users started using it, data came in, and Gentoo learned from that data to provide better recommendations, after going through this cycle, we continuously gained confidence that if these two hypotheses are correct, we can truly create a world where shopping is possible through conversation, and that we can innovate the user experience within the shopping vertical with conversation, the most intuitive interface.

Youngmoo Kim: WADDLE has been doing e-commerce AI for quite a long time, even long before LLMs like ChatGPT came out, and they've been continuously building up experience, networks, know-how, and technologies with an unwavering belief that the AI era would come. That's why they were able to move faster than anyone else after ChatGPT came out, thinking about how to further utilize it.

And there's one more thing that impressed me. When we first did IR, WADDLE said that our ultimate competitor is not e-commerce AI, but Siri. I found that part very impressive because many people used to say, 'Won't Siri just monopolize everything if chatbots and voice AI take over someday?' Could you explain how WADDLE plans to compete with Siri in the future?

For Gentoo, the real competition is Siri.

Yongwon Cho: Not just Siri, but AGI or any general-purpose AI assistant represented by Siri is still considered our biggest long-term competitor. The pace of AI technology evolution is always years faster than I anticipate. Technologies like OpenAI's Agents SDK, Anthropic's MCP, and Gemini's A2A are ultimately crucial stepping stones in the process of creating Artificial General Intelligence (AGI). So, I've been thinking lately that our long-term competitors might not be so long-term after all.

However, I believe that an AI agent that has best designed the user experience in a vertical domain can provide utility for a not-so-short period until AGI develops. Our current strategy is to create an agent that innovates the user experience in the shopping domain.

At the same time, I believe the day will come, not in too distant future, when Gentoo will not only exist within shopping malls but also function as a personal shopping assistant at the consumer level. That's why we are tracking the monthly or weekly number of users conversing with Gentoo as a very important metric. We are focusing on becoming the best AI agent in the shopping domain, to the extent that we are monitoring how many of the 50 million people in Korea are becoming accustomed to shopping through conversation.

Youngmoo Kim: I was very impressed by WADDLE's vision of quickly catching up with market changes in a short period, leading the AI agent scene, especially in e-commerce, within Korea, and eventually evolving into a company that can take over all AI-based interfaces in the distant future.

I have high expectations for its future growth. We will conclude the interview here. Thank you for listening to the interview with WADDLE. Do you have any final words?

Yongwon Cho: WADDLE is, to my knowledge, the best and only team in Korea that is innovating the e-commerce customer experience using AI agents and conversational interfaces. We hope you will continue to follow the innovation in customer experience through AI that we will create.

I believe AI is no longer solely the realm of R&D for scientists, but has reached the stage of commercialization by engineers and businesses. Please continue to support and look forward to WADDLE, who is excelling by being the first to penetrate the market. Thank you.